Production log · Limited firsthand observations
How we made our first AI episode
The actual production path behind Enter the Forge: what stalled, what we reused, and how the finished video was assembled.
The result: one finished episode, several lessons.
We published a 44-second vertical episode using five selected voice recordings, approved artwork, captions and an original electronic score. The completed cut uses camera movement and illustrated animation. It does not have full character lip-sync.
Also watch on TikTok or Facebook.
What we were trying to make
Enter the Forge introduces a fictional world where AI claims face an evidence check. We needed recurring character designs and distinct voices, so replacing approved assets with whatever an automatic generator selected was not an acceptable finish.
Where the first attempts stopped
The InVideo attempt ran into its available credit limit without producing the finished episode. In ElevenLabs Studio, we uploaded five MP3 recordings and five reference images. Batch analysis failed, smaller batches failed, and an individual transcription attempt also failed. We stopped the repeated attempts; the agent displayed 356 credits used. That figure is the displayed agent total, not a complete project cost or a dollar charge.
These observations describe one production session. They do not establish either product's general failure rate, and we have not determined the underlying cause of the Studio errors.
What finally worked
- Keep the selected voice recordings and original artwork as separate files.
- Place the dialogue on a timeline without changing voice speed.
- Build shots around the spoken lines, with readable captions and restrained motion.
- Render locally using Python and FFmpeg, then inspect representative frames.
- Correct the character framing, hold the opening message for four seconds, and export a 1080 × 1920 MP4.
The owner supplied the ElevenLabs voice files. The final edit was assembled locally; it was not a finished export from InVideo or ElevenLabs Studio. The separate villain monologue was not used.
Checks and limits
The final 44-second H.264/AAC file decoded successfully. Audio measurement was approximately −16.7 LUFS integrated with a −1.1 dBTP true peak. Those are technical checks, not proof that every line sounds natural. The assistant did not conduct a listening review, and the record does not establish the commercial entitlement of each supplied recording. A production workflow should include a human listening pass and a check of the license that applies to its audio.
TikTok accepted the post after review. Facebook's copyright check reported no issues and the Reel was published publicly. A platform check is not a substitute for checking asset rights. No audience-growth or book-sales result is claimed here.
The lesson for your next video
Test one complete scene before requesting an entire episode. Preserve anything you have already approved. Budget for revisions, and keep an editor-based route available when generation or analysis stalls. A finished, inspectable cut gives you a better basis for improvement than repeated attempts at a perfect one-shot result.
Explore the tools
Compare voice workflows · Compare video workflows · Read the CapCut research review
Episode dialogue transcript
ForgeFather: Welcome to the Forge. Around here, hype doesn't survive without evidence.
Challenger: I am the most advanced AI ever.
ForgeFather: Claims mean nothing until they face the Forge.
Forge Bot: Scanning claims, user feedback, limitations, pricing, and evidence.
Glitch: Wait… they're checking the truth behind the hype.
System: Warning. Unknown AI detected.
Script transcript; recordings were not independently transcribed for this log.